An architecture diagram with only a user, a model and a database tells me almost nothing about whether the system is ready for production. I want to see who authenticated the user, what evidence they were allowed to retrieve, which actions were permitted, which credential executed them and what happens when a dependency fails.
Northstar's application authenticates the user and establishes the tenant and permitted task. The controller selects context, invokes the model and evaluates proposed actions. Retrieval releases only authorized evidence. Tool services execute business operations with scoped credentials. Validation determines how the resulting output may be displayed or used downstream.
A model gateway can centralize provider selection, usage limits and approved routes. It does not replace record authorization. A fallback provider should be eligible only when it meets the same data restrictions and passes the evaluation required for the task.
Source ingestion sits outside the live request path, but it still changes live behaviour. Ingestion validates sources and maintains versions and access metadata. If freshness, deletion or permissions drift there, answers can become wrong without any model setting changing.
Monitoring should connect a request to the versions involved, evidence identifiers, tool outcomes and operational measurements. Avoid routinely recording raw personal data or secrets. Protect debugging captures with explicit access and retention controls. A trace should help diagnose an event without creating a second disclosure path.
Do not turn every box in the architecture into a separate service by default. A small application can combine several responsibilities cleanly. Split them when isolation, scale, ownership or failure containment justifies the operating cost.